Skip to content

请教参数设定 #44

Description

@AragornHorse

我在8xH100上做的测试,用的Qwen2.5-7B-base,1个卡放reference model,4个卡rollout,3个卡训练。
由于原始参数下会OOM,下面是我现在的主要参数设定:
beta = 0.04
all_steps = 1000
Q_batch_size = 4
num_pre_Q = 8
train_batch_size = 2
gen_update_steps = 16
save_steps = 200
compute_gen_logps = True
clip_param = 0.2
gradient_accumulation_steps = 4
lr = 1e-6
max_tokens = 700

我花了约1个小时使用gsm训练了800步,gsm用evalscope上测试的准确率只从0.08涨到了0.14,没有达到这里Readme中的效果,请问是因为我的参数设置问题吗?如果是的话有没有办法优化。

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions